Measuring The Progress Of Neural Machine Translation:A Combined Approach Of Diagnostic Evaluation And Ranking

Machine translation has long been a focus of study within the discipline of computational linguistics,but no major breakthroughs have been made for decades.However,spurred by the recent advances in artificial intelligence,a new approach to Machine Translation,Neural Machine Translation(NMT),became known to the public.Particularly,the NMT system developed by Google(GNMT)received international attention.According to the official report,GNMT is capable of improving translation quality by as much as 55%to 85%compared to conventional MT systems,giving rise to two widely accepted beliefs.First,NMT systems like GNMT will soon render human translators obsolete;second,GNMT represents the state-of-the-art for NMT platforms,and is therefore superior to any other NMT systems.To measure the current progress of the NMT technology and put the two claims to test,this paper conducts a test run with domain-specific materials on three commonly used NMT systems:GNMT,Baidu Translate and Youdao NMT(YNMT).It then adopts a combined approach of diagnostic evaluation and ranking to assess the translation quality of each candidate system.Statistics from the study reveal two facts.First,despite achieving remarkable improvements compared to conventional MT paradigms,NMT systems are still far from matching human-level translation.Also,the performances of NMT systems vary greatly depending on the source language and material domains.Therefore,while GNMT only manages to give a stably acceptable performance,home-based Baidu Translate and YNMT produces markedly better results by comparison.This study is meaningful to translation training and learning in that it not only provides a technologically accessible portrayal of NMT's current state,but also touches upon the impact such technology has on the future of human translators.